Deadlock workaround

[mbrandonw]

Right now we do a DispatchQueue.main.sync in order to make sure that the test observer is registered on the main thread. sync is always a dangerous operation, and it turns out we can deadlock. It can be reproduced with the following 3 steps:

1. Start background thread/queue and access a dependency after a small amount of time.
2. Before the dependency is accessed block the main thread with some work.
3. After the main thread work is done access another dependency.

This causes a deadlock because the background thread has already entered into a dispatch_once to resolve static var current but is currently waiting on the main thread to finish its work. And simultaneously the main thread has now also entered the dispatch_once, and hence a deadlock.

The fix is perhaps a little risky, but we think it should be ok. We are going to allow adding the test observer in main.async if the dependency is first being accessed on a background thread. It seems theoretically possible that the test observer will not be registered by the time the test ends, but we haven't been able to find a situation in which that occurs.

[mbrandonw]

Incidentally this will fix #76 too.

[stephencelis]

Looks good!